A lightweight weed detection model for cotton fields based on an improved YOLOv8n

Area of Science:

• Agricultural Science
• Computer Vision
• Deep Learning

Background:

• Weed proliferation in cotton fields significantly impacts crop yield and health.
• Deep learning models offer high-precision weed recognition but often suffer from high computational demands and resource consumption.
• There is a critical need for efficient and lightweight weed detection methods for practical implementation in agriculture.

Purpose of the Study:

• To develop an efficient and lightweight deep learning algorithm for detecting weeds in cotton fields.
• To address the limitations of existing complex deep learning models in terms of computational cost and resource usage.
• To enhance the practical applicability of automated weed recognition systems in agriculture.

Main Methods:

• Proposed the YOLO-Weed Nano algorithm, an optimized version of the YOLOv8n model.
• Integrated Depthwise Separable Convolution (DSC) into the HGNetV2 network (DS_HGNetV2) as the model backbone.
• Incorporated Bidirectional Feature Pyramid Network (BiFPN) for enhanced feature fusion and a lightweight LiteDetect head for reduced computation.

Main Results:

• YOLO-Weed Nano demonstrated a 1% improvement in mean Average Precision (mAP) compared to the original YOLOv8n model.
• The proposed model achieved significant reductions in parameters (63.8%), computation (42%), and weights (60.7%).
• These optimizations make the model more suitable for deployment in resource-constrained agricultural environments.

Conclusions:

• The YOLO-Weed Nano algorithm presents a highly efficient and lightweight solution for cotton field weed detection.
• The model successfully balances high detection accuracy with reduced computational complexity.
• This advancement facilitates the practical application of deep learning for effective weed management in modern agriculture.